MSCR: Jointly Balancing Modality Utilization and Discovering Synergistic Information

In multimodal learning, modality imbalance often causes one modality to dominate the optimization process, preventing the model from fully exploiting cross-modal interactions. Recent balance-oriented methods, such as Multimodal Competition Regularizer (MCR), have achieved strong performance by alleviating modality imbalance during training. However, existing dynamic balancing methods mainly focus on adjusting modality contributions and do not explicitly consider how modality imbalance affects the formation of synergistic information in multimodal fusion. Yet such synergy is a key source of multimodal advantage, as it reflects predictive gain that cannot be recovered from unimodal evidence alone. To address this gap, we propose Multimodal Synergistic Competition Regularizer (MSCR), a multimodal fusion method that jointly promotes balanced modality utilization and synergistic information discovery. MSCR contains two complementary objectives: an anti-dominance objective that suppresses excessive unimodal influence and creates room for synergistic information discovery, and a synergy objective that explicitly encourages the fused predictor to surpass a non-synergistic unimodal reference. We further adopt an adaptive modulation strategy to emphasize synergy enhancement on samples that require stronger collaborative modeling. Experiments show that MSCR consistently improves performance, enhances measured synergistic information, and remains effective under transformer-based fusion backbones.